Creating a Dead Poets Society:
Extracting a Social Network of Historical Persons
from the Web
Gijs Geleijnse
Jan Korst
Philips Research
High Tech Campus 34, 5656 AE Eindhoven, The Netherlands
{gijs.geleijnse,jan.korst}@philips.com
Abstract. We present a simple method to extract information from search engine
snippets. Although the techniques presented are domain independent, this work
focuses on extracting biographical information of historical persons from multiple unstructured sources on the Web. We first similarly find a list of persons and
their periods of life by querying the periods and scanning the retrieved snippets
for person names. Subsequently, we find biographical information for the persons
extracted. In order to get insight in the mutual relations among the persons identified, we create a social network using co-occurrences on the Web. Although we
use uncontrolled and unstructured Web sources, the information extracted is reliable. Moreover we show that Web Information Extraction can be used to create
both informative and enjoyable applications.
1 Introduction
Information extraction (IE) is the task of identifying instances and relations
between those instances in a text corpus. Thanks to popular search engines,
the Web is a well-indexed tera-size corpus, where precise queries often lead to
highly relevant text fragments. When we assume that all knowledge available on
a domain can be found on the Web, this corpus can be used for ontology-driven
information extraction [1], i.e. we can assume that all required information is
available within the web. Due to the redundancy of information on the web, the
question is rather how to identify at least one formulation of every fact of interest (determined by the ontology), than how to extract every relevant formulation
within the corpus (i.e. corpus-driven information extraction).
Pattern-based approaches have shown to be effective for ontology-driven
information extraction. Such approaches can not only be used to identify lexicographical relations (e.g. hyponyms, part-of) [2, 3] but also various other relations. Patterns expressing relations (e.g. ‘is the president of’, [2]) or combined
instance-pattern terms (e.g. is the president of the United States, [4, 1] are used
as queries to a search engine. Such queries give access to highly relevant text
fragments and instances (e.g. George W. Bush) and the expressed relations can
be extracted simultaneously.
In this paper we show how web information extraction can indeed be a
valuable addition to the current collective knowledge on the web. We present
a method to populate an ontology using information extraction from search engine query results. Using the redundancy of information on the web, we can
apply simple methods to extract term and relations. We illustrate this approach
by populating an ontology with historical persons, their main biographical data
and their ‘degree of fame’. We also create a social network among the persons
extracted. By combining and structuring information from a large collection of
web pages, questions like Who are the most famous novelists from Ireland?,
Which people were born in 1648?, Who are popular female composers? and
Who are considered to be most related to Vincent van Gogh? can now be easily
answered. This work illustrates that using simple techniques we can create a
reliable ontology that gives insight in the domain and can be a true addition to
the information available.
2 Related Work
KnowItAll is a hybrid named-entity extraction system [2] that finds lists of instances (e.g. ’Busan’, ’Karlsruhe’)of a given class (e.g. ’City’) from the web
using a search engine. It combines hyponym patterns [5] and class-specific,
learned patterns to identify and extract named-entities. Moreover, it uses adaptive wrapper algorithms [6] to extract information from HTML markup such as
tables. Contrary to the method used in this paper, the instances found are not
used to create new queries. In [7] the information extracted by KnowItAll is
evaluated using a combinatorial model based on the redundancy of information
on the web.
Recently, the KnowItAll project has addressed the identification of complex
named entities (such as book titles), using a statistical n-gram approach [8].
In [9, 4] such complex entities are recognized using a set of simple rules. For
example, a movie title is recognized when it is placed between quotation marks.
Cimiano and Staab [10] describe a method to use a search engine to verify a
hypothesis relation. For example, the number of hits to “rivers such as the Nile”
and “cities such as the Nile”) are compared to determine the class of Nile. Per
instance, the number of queries is linear in the number of classes considered.
Where Cimiano and Staab assume the relation patterns to be given, in [11]
a technique is introduced to find precise patterns using a training set of related
instances. In [4] a method is presented to extract information from the web using
effective patterns, as precision is not the only criterion for a pattern to provide
useful results.
The number of search engine hits for pairs of instances can be used to compute a semantic distance between the instances [12]. The nature of the relation
is not identified, but the technique can for example be used to cluster related
instances. In [13] a similar method is used to cluster artists using search engine
counts. However, the total number of hits provided by the search engine is an estimate and not always reliable [14]. In [15] an approach is presented where one
instance is queried and the resulting texts are mined for occurrences of other
instances. Such an approach is not only more efficient in the number of queries,
but also gives better results.
The extraction of social networks using web data is a frequently addressed
topic. For example, Mori et al. [16] use tf·idf to identify relations between politicians and locations and [17] use inner-sentence co-occurrences of company
names to identify a network of related companies.
3 Problem Description
Suppose that we are given an ontology O with n classes c1 , ..., cn and a set
R of relation classes Ri,j on classes ci and cj . For example, if ‘Person’ and
‘City’ are classes, then ‘is born in(Person, City)’ is a relation class on these
classes. The tuple ‘(Napoleon Bonaparte, Ajaccio)’ may be an instance of ‘is
born in(Person, City)’ if ‘Napoleon Bonaparte’ is an instance of ‘Person’ and
‘Ajaccio’ is an instance of ‘City’. We call a class complete if all instances are
assumed to be given beforehand. For empty classes no instances are given.
Problem. Given the ontology O,
(1.) populate the incomplete classes with instances extracted from the web, and
(2.) populate the relation classes in R by identifying formulations of all relations
Ri,j between the instances of ci and cj .
¤
In this work we focus on the population of an ontology on historical persons. In the given ontology (cf. Figure 1) all classes but Person are complete,
while Person is empty. The class Period contains all combinations of years that
are likely to denote the periods of life of a historical person. For example ’1345
- 1397’ is an instance of Period. The class Nationality contains all names of
countries in the world. We identify derivatives of country names as well and use
them as synonyms (e.g. American for United States, English for United Kingdom and Flemish for Belgium). A hierarchy among the instances of Nationality
is defined using the names of the continent, such that we can for example select
a list of historical persons from Europe. Likewise, the instances of Profession
reflect 88 professions. For the instances male and female we have added a list
Person
has
Nationality
Profession
has
has
Gender
related_with
has
lived
Fame
Period
Fig. 1. The ontology on historical persons to be populated.
of derivatives to be used as synonyms, namely the terms he, his, son of, brother
of, father of, man and men for male and the analogous words for female. We
use the class Fame to rank the retrieved instances of Person according to their
presence on the web. Hence the task is to identify a collection of biographies of
historical persons and to identify a social network between the persons found.
As persons may have more than one profession and can be related to multiple
other people, we are interested in a ranked list of professions and related persons
for each historical person found.
For efficiency reasons, we only extract information from the snippets returned by the search engine. Hence, we do not download full documents. As
only a limited amount of automated queries per day are allowed1 , the approach
should be efficient in the number of queries.
4 Finding Biographies
We use the given instances in the ontology O to populate the class Person, i.e.
to find a list of names of historical persons. We use the instances available in
the given ontology to formulate queries and hence create a corpus to extract
information from.
Using a pattern-based approach, we combine natural language formulations
of relations in O with known instances into queries. Such queries lead to highly
relevant search results, as the snippets are expected to contain related instances.
As search engines return only a limited amount of results and the same names
may occur in multiple search results, we have to identify effective queries [4].
Suppose we use instances in the class Profession to extract the persons.
When querying for the instance ’composer’, it is likely that few well-known
composers dominate the search results. As we are interested in a rich ontology
of historical persons, this is thus a less-suited approach.
The class Period contains all combinations of years that are likely to denote
the periods of life of a historical person. Hence, the number of instances known
1
Yahoo! accepts 5,000 queries per day.
for the class Period is by far the largest for all complete classes in O. As it is
unlikely that many important historical persons share both a date of birth and a
date of death, the use of this class is best suited to obtain a long and diverse list
of persons. The names of historical persons are often preceded in texts with a
period in years (e.g. ‘Vincent van Gogh (1853 - 1890)’). As this period is likely
to denote the period he or she lived in, we choose the pattern ”(year of birth –
year of death)” to collect snippets to identify the names of historical persons.
4.1 Identifying person names in snippets
Having obtained a collection of snippets, the next problem is to extract instances
from the texts, in this case persons names. We choose to identify the names
within the snippets using a rule-based approach. Although the design of rules is
laborious, we do not opt for an approach based on machine learning (e.g. [18,
19, 8]) for the following reasons.
1. No representative training set is available. The corpus – the collection of
snippets – contains broken sentences and improper formulations. Moreover,
the corpus is multi-lingual.
2. The Named Entity Recognition (NER) task is simplified by the use of the
patterns. Since we expect person names preceding the period queried, we
know the placeholder of the named entity (i.e. preceding the queried expression). In a general NER task this information is not available.
3. Since the corpus consists of uncontrolled texts, there is need for post-processing. The snippets may contain typos, missing first or last names, and
errors in dates and names.
We hence use a rule-based approach to identify person names in the snippets
found with the periods.
First we extract all terms directly preceding the queried expressing that
match a regular expression. That is, we extract terms of two or three capitalized
words and compensate for initials, inversions (e.g. ’Bach, Johann Sebastian’),
middle names, Scottish names (e.g. McCloud) and the like.
Subsequently, we remove extracted terms that contain a word in a tabu list
(e.g. ‘Biography’) and names that only occur once within the snippets. Having
filtered out a set of potential names of persons, we use a string matching among
the extracted names to remove typos and names extracted for the wrong period.
Using the 80,665 periods identified, we obtain a list of 28,862 terms to be
added as instance to the class Person. Simultaneously, we extract the relations
between the periods queried and the extracted instances.
Napoleon Bonaparte was the greatest military genius of the 19th century
Napoleon Bonaparte was born of lower noble status in Ajaccio, Corsica on August 15,
1769
Napoleon Bonaparte was effectively dictator of France beginning in 1799 and
Napoleon Bonaparte was the emperor of France in the early 1800s
Napoleon Bonaparte was a bully, rude and insulting
Napoleon Bonaparte was in Egypt and was not enjoying his tour
Napoleon Bonaparte was a great warrior and a notable conqueror
Napoleon Bonaparte was born on August 15, 1769 to Carlo and Letizia Bonaparte
Napoleon Bonaparte was defeated at Waterloo
Table 1. Example search results for the query ’Napoleon Bonaparte was’.
In the evaluation section we analyze the quality of the extracted instances
and compare the rule-based approach with a state-of-the-art named entity recognizer based on machine learning.
4.2 Using Mined Names to find additional biographical information
Having found a list of instances of the class Person, we first determine a ranking
of the instances extracted.
Finding a Rank. To present the extracted information in an entertaining manner, we determined the number of hits for each identified person. As names are
not always unique descriptors, we queried for the combination of the last name
and period (e.g. ’Rubens (1577 - 1640)’). Although the number of hits returned
a search engine is an estimate and irregularities may occur [14], we consider
this simple and efficient technique to be well suited for this purpose.
Now we use the names of these instances in a similar fashion to acquire biographical information for the 10,000 best ranked persons. To limit the number of
queries per instance, we select the pattern ’was’ to reflect the relation between
Person on the one hand and Nationality, Gender and Profession on the other
hand. By querying phrases such as ’Napoleon Bonaparte was’ we thus expect
to acquire sentences containing the biographical information. Table 1 contains
examples of the sentences used to determine biographical information. We scan
these sentences for occurrences of the instances (and their synonyms) of the
related classes.
Relating persons to a gender. We simply counted instances and their synonyms
within the snippets that refer to the gender ‘male’ the opposite words that refer
‘female’. We simply related each instance of Person to the gender with the highest count.
Relating persons to a nationality. We assigned the nationality with the highest
count.
Relating persons to professions. For each person, we assigned the profession
p that most frequently occurred within the snippets retrieved. Moreover, as persons may have multiple professions, all other professions with a count at least
half of the count of p were added.
Hence, using one query per instance of Person, we identify basic biographical information.
5 Identifying a Social Network
Having gathered a list of historical persons with biographical information, we
are interested how the persons in the list are perceived to be related. Obviously
such information can be extracted from the biographies, e.g. persons can be
considered related when they share a profession, have the same nationality or
lived in the same period.
However, we are interested in the way people nowadays relate the historical
people extracted. For example, we are interested to identify the person who
is considered to be most related to Winston Churchill. We therefore mine the
Web for a social network of people extracted using the method in the previous
section.
We assume that two persons are related when they are often mentioned in the
same context. In early work, search engine queries were used that contained both
the terms (e.g. [10, 12]). The number of Google hits to such queries were used as
co-occurrence count co(u, v). However, work in [15] has shown that a patternbased approach is not only more efficient in the number of queries, but also gives
better results. Hence, we opt for an approach similar as described in the previous
section, where we use patterns to identify relatedness between persons. We use
the total numbers of co-occurrences co(p, q) for persons p 6= q to compute the
relatedness T (p, q) of p to q. Using the hypothesis that enumerated items are
often related, we use patterns expressing enumerations (Table 2) to obtain the
snippets.
For each historical person p we could consider the person q with the highest
co(p, q) to be the most related to p. However, we observe that, in that case,
frequently occurring persons have a relatively large probability to be related to
any person. This observation leads to a normalized approach.
co(a, b)
c,c6=b co(c, b)
T (a, b) = P
(1)
For each of the best 3,000 ranked persons, we computed a ranked list of
most related persons in the large set of the 10,000 persons with biographies.
”like [person] and [person]”
”such as [person] and [person]”
”including [person] and [person]”
”for example [person] and [person]”
”namely [person] and [person]”
”[person] and [person]”
”[person] [person] and other”
Table 2. Patterns used to find co-occurrences within the search engine snippets.
6 Experimental Results
In this section we discuss the results of the ontology population method applied
to the ontology on historical persons. We present examples of the extracted data
to give the reader an impression of the results. Moreover, we show that structured data can be used to gain insights. Section 6.1 handles the extracted instances of Person and the identified biographical relations, while Section 6.2
handles the social network identified of the extracted persons. As no prior work
is known in this domain, we cannot compare the results with others.
6.1 Evaluating the identified biographical information
The rank assigned to each of the persons in the list provides a mechanism to
present the extracted data in an attractive manner. Table 3 gives the list of the
25 best ranked persons and the identified biographical information. Using the
criterion defined in Section 4, Johann Sebastian Bach is thus the best known
historical figure.
As the data is structured, we can also perform queries to select subsets of
the full ranked list of persons. For example, we can create a list of best ranked
artists (Table 4), or a ‘society’ of poets (Table 5). We note that Frédéric Chopin
is often referred to as ’the poet of the piano’. Table 6 shows that Vincent van
Gogh is the best ranked Dutch painter.
The reader can verify that the given list of extracted persons are highly accurate. However, lacking a benchmark set of the best known historical persons,
we manually evaluated samples of the extracted ontology to estimate precision
and recall.
Precision. To estimate the precision of the class Person, we selected three
decennia, namely 1220-1229, 1550-1559 and 1880-1889, and analyzed for each
the candidate persons that were found to be born in this decennium. For the
first two decennia we analyzed the complete list, for decennium 1880-1889 we
analyzed only the first 1000 as well as the last 1000 names. This resulted in
a precision of 0.94, 0.95, and 0.98, respectively. As the decennium of 18801889 resulted in considerably more names, we take a weighted average of these
results. This yields an estimated precision for the complete list of 0.98.
Johann Sebastian Bach (1685-1750)
Wolfgang Amadeus Mozart (1756-1791)
Ludwig van Beethoven (1770-1827)
Albert Einstein (1879-1955)
Franz Schubert (1797-1828)
Johannes Brahms (1833-1897)
William Shakespeare (1564-1616)
Joseph Haydn (1732-1809)
Johann Wolfgang Goethe (1749-1832)
Charles Darwin (1809-1882)
Robert Schumann (1810-1856)
Leonardo da Vinci (1452-1519)
Giuseppe Verdi (1813-1901)
Frederic Chopin (1810-1849)
Antonio Vivaldi (1678-1741)
Richard Wagner (1813-1883)
Ronald Reagan (1911-2004)
Franz Liszt (1811-1886)
Claude Debussy (1862-1918)
Henry Purcell (1659-1695)
Immanuel Kant (1724-1804)
James Joyce (1882-1941)
Friedrich Schiller (1759-1805)
Georg Philipp Telemann (1681-1767)
Antonin Dvorak (1841-1904)
Germany
Austria
Germany
Germany
Austria
Germany
United Kingdom
Austria
Germany
United Kingdom
Germany
Italy
Italy
Poland
Italy
Germany
United States
Hungary
France
United Kingdom
Germany
Ireland
Germany
Germany
Czech Republic
composer,organist
composer,musician
composer
scientist,physicist
composer
composer
author,poet
composer
philosopher,director,poet..
naturalist
composer
artist,scientist,inventor
composer
composer,pianist,poet
composer
composer
president
pianist,composer
composer
composer
philosopher
author
poet,dramatist
composer
composer
Table 3. The 25 historical persons with the highest rank.
We compare the precision of the rule-based approach with a state-of-theart machine-learning-based algorithm, the Stanford Named Entity Recognizer
(SNER [20]), trained on the CoNLL 2003 English training data. Focussing on
persons born in the year 1882, using the rule-based approach we extracted 1,211
terms. SNER identified 24,652 unique terms as person names in the same snippets. When we apply the same post-processing on SNER extracted data (i.e.
removing typos by string matching, single-word names and names extracted for
different periods), 2,760 terms remain, of which 842 overlap with the terms extracted using the rule-based approach.
We manually inspected each of these 2,760 terms, resulting in a precision of
only 62%. Around half of the correctly extracted names are not recognized by
the rule-based approach, most of them due to the fact that these names did not
directly preceded the queried period.
To estimate the precision of the extracted biographical relations, we inspected randomly selected sublists of the top 2500 persons. When we focus
on the best scoring professions for the 2500 persons, we estimate the precision
of this relation to be 96%. We did not encounter erroneously assigned genders,
while we found 98% of the cases the right Nationality, if one is found.
Leonardo da Vinci (1452 - 1519)
Pablo Picasso (1881 - 1973)
Vincent van Gogh (1853 - 1890)
Claude Monet (1840 - 1926)
Pierre-Auguste Renoir (1841 - 1919)
Paul Gauguin (1848 - 1903)
Edgar Degas (1834 - 1917)
Paul Cezanne (1839 - 1906)
Salvador Dali (1904 - 1989)
Henri Michaux (1899 - 1984)
Gustav Klimt (1862 - 1918)
Peter Paul Rubens (1577 - 1640)
Katsushika Hokusai (1760 - 1849)
Amedeo Modigliani (1884 - 1920)
JMW Turner (1775 - 1851)
James Mcneill Whistler (1834 - 1903)
Rene Magritte (1898 - 1967)
Henri Matisse (1869 - 1954)
Rembrandt van Rijn (1606 - 1669)
Edouard Manet (1832 - 1883)
Herm Albright (1876 - 1944)
Marc Chagall (1887 - 1985)
Edvard Munch (1863 - 1944)
Wassily Kandinsky (1866 - 1944)
Francisco Goya (1746 - 1828)
Italy
Spain
Netherlands
France
France
France
France
France
Spain
Belgium
Austria
Belgium
Japan
Italy
United Kingdom
United States
Belgium
France
Netherlands
France
Russia
Norway
Russia
Spain
artist, scientist,...
artist
artist, painter
artist, painter,...
painter
painter
artist, painter,...
painter, artist
artist
artist, poet
painter, artist
artist, painter
painter
artist, painter
artist, painter
artist
artist, painter
artist
artist, painter
artist, painter
artist, engraver,...
painter, artist
painter, artist
artist, painter
artist, painter
Table 4. The 25 artists with the highest rank.
Hence, we conclude that the ontology populated using the rule-based approach is precise.
Recall. We estimate the recall of the instances found for Person by choosing
a diverse set of six books containing short biographies of historical persons. Of
the 1049 persons named in the books, 1033 were present in our list, which gives
a recall of 0.98. For further details on the chosen books we refer to [21].
From Wikipedia, we extracted a list of important 1882-born people2 . The
list contains 44 persons. Of these 44 persons, 34 are indeed mentioned in the
Google snippets found with the queried patterns. Using the rule-based approach,
we identified 24 of these persons within the snippets. The other ones were only
mentioned once (and hence not recognized) or found in different places in the
snippets, i.e. not directly preceding the queried period. Using SNER, we identified 27 persons from the Wikipedia list.
For the recall of the identified biographical relations, we observe that for the
10,000 persons that we considered all were given a gender, 77% were given a
nationality, and 95% were given one or more professions.
2
http://en.wikipedia.org/wiki/1882, visited January 2007
William Shakespeare (1564-1616)
Johann Wolfgang Goethe (1749-1832)
Frederic Chopin (1810-1849)
Friedrich Schiller (1759-1805)
Oscar Wilde (1854-1900)
Jorge Luis Borges (1899-1986)
Victor Hugo (1802-1885)
Ralph Waldo Emerson (1803-1882)
William Blake (1757-1827)
Dante Alighieri (1265-1321)
Robert Frost (1874-1963)
Heinrich Heine (1797-1856)
Robert Louis Stevenson (1850-1894)
Alexander Pope (1688-1744)
Hildegard von Bingen (1098-1179)
Lord Byron (1788-1824)
John Donne (1572-1631)
Henri Michaux (1899-1984)
Walt Whitman (1819-1892)
Robert Burns (1759-1796)
United Kingdom
Germany
Poland
Germany
Ireland
Argentina
France
United States
United Kingdom
Italy
United States
Germany
Samoa
United Kingdom
Germany
Greece
United Kingdom
Belgium
United States
United Kingdom
author,poet
poet, psychologist, philosopher..
composer,pianist,poet
poet,dramatist
author,poet
author,poet
author,poet,novelist
poet,philosopher,author
poet
poet
poet
poet
engineer,author,poet
poet
composer,scientist,poet
poet
poet,author
poet
poet
poet
Table 5. The 20 best ranked poets.
Vincent van Gogh (1853-1890)
Rembrandt van Rijn (1606-1669)
Johannes Vermeer (1632-1675)
Piet Mondrian (1872-1944)
Carel Fabritius (1622-1654)
Kees van Dongen (1877-1968)
Willem de Kooning (1904-1997)
Pieter de Hooch (1629-1684)
Jan Steen (1626-1679)
Adriaen van Ostade (1610-1685)
Table 6. The 10 best ranked painters from the Netherlands.
6.2 Evaluating the social network
Aiming for a reflection of the collective knowledge of web contributors on historical figures, the extracted social network of historical persons is not a verifiable collection of facts. We illustrate the social network extracted by two examples. Figure 3 depicts the relatedness among the best ranked persons. An arrow
from person p to q is drawn if q is among the 20 nearest neighbors of p. Using the same criterion, Figure 4 depicts the relatedness among the best ranked
authors.
We are able to verify the precision of the relatedness between historical
persons if we make the following assumptions. We consider two persons to be
related if either (1.) they lived in the same period, i.e. there is an overlap in the
periods the two lived, (2.) they shared a profession, (3.) they shared a nationality,
or (4.) they are both female.
precision of the social network
1
0.98
0.96
0.94
0.6
0.92
precision
precision
precision of the social network per criterion for 10 Nearest Neighbors
0.9
same nationality
same profession
0.8
same period
0.7
both female
1-NN
5-NN
10-NN
15-NN
25-NN
0.9
0.88
0.5
0.4
0.3
0.86
0.84
0.2
0.82
0.1
0.8
0
0
200 400 600 800 1000 1200 1400 1600 1800 2000
n
0
200 400 600 800 1000 1200 1400 1600 1800 2000
n
Fig. 2. (l.) Precision for the social network for the n best ranked persons and their k nearest
neighbors. (r.) Precision for 10-NN per criterion.
Robert Schumann
Johannes Brahms
Ludwig van Beethoven
Johann Sebastian Bach
Franz Schubert
Wolfgang Amadeus Mozart
Frederic Chopin
Giuseppe Verdi
Leonardo da Vinci
Albert Einstein
Johann Wolfgang Goethe
William Shakespeare
Joseph Haydn
Charles Darwin
Fig. 3. The extracted social network for the 15 most best ranked persons.
Of course we cannot evaluate recall of the algorithm on these criteria, as for
example not all persons sharing a nationality need to be consider to be related.
We therefore the evaluate precision of the social network on these “minimal criteria”. For the 3,000 best ranked persons, we select the k most related persons.
Per pair we evaluate whether either one of the four criteria is being met. This
precision rate is presented in Figure 2 (l.). In comparison, the probability of any
of the 3,000 persons to be related to a person in the large list of 10,000 is 45%.
The precision rates for the social network per criterion can be found in Figure 2
(r.). The probabilities for two randomly selected persons to share a period, profession and nationality are 38%, 7.5% and 6.5% respectively. The chance that
two historical persons are both female is only 0.5%. We hence conclude that
these results give good confidence on the quality of extracted social network.
7 Conclusions
We illustrated a simple method to populate an ontology using a web search
engine by finding biographical information on historical persons. Starting with
Robertson Davies
Johann Wolfgang Goethe
Friedrich Schiller
HL Mencken
Charles Dickens
Mark Twain
Oscar Wilde
Washington Irving
Prentice Mulford
William Blake
William Shakespeare
Heinrich Heine
Robert Frost
Francis Bacon
James Joyce
Jorge Luis Borges
George Bernard Shaw
Victor Hugo
Ralph Waldo Emerson
Henry David Thoreau
Dante Alighieri
Fig. 4. The extracted social network for the best ranked authors.
the empty class Person, we found over 28 thousand historical persons with a
high precision rate. The biographical information identified for the best ranked
persons has shown to be of high quality as well. The same method is used to
create a social network for the historical persons, with convincing results.
Hence, we show that simple Web Information Extraction techniques can
be used to precisely populate ontologies. By combining and structuring information from the Web, we create a valuable surplus to the knowledge already
available.
In future work, we plan to further address the automatic identification of
characterizations for other items such as movies and musical artists. The identification of collective knowledge and opinions is perhaps more interesting than
collecting plain facts, which often can be mined from semi-structured sources.
By combining information extracted from multiple web pages, we plan to research methods to automatically tag items.
References
1. McDowell, L., Cafarella, M.J.: Ontology-driven information extraction with ontosyphon. In:
Proceedings of the 5th International Semantic Web Conference (ISWC 2006). Volume 4273
of LNCS., Athens, GA, Springer (2006) 428 – 444
2. Etzioni, O., Cafarella, M.J., Downey, D., Popescu, A., Shaked, T., Soderland, S., Weld, D.S.,
Yates, A.: Unsupervised named-entity extraction from the web: An experimental study.
Artificial Intelligence 165(1) (2005) 91 – 134
3. van Hage, W.R., Kolb, H., Schreiber, G.: A method for learning part-whole relations. In:
Proceedings of the 5th International Semantic Web Conference (ISWC 2006). Volume 4273
of LNCS., Athens, GA, Springer (2006) 723 – 736
4. Geleijnse, G., Korst, J.: Learning effective surface text patterns for information extraction. In: Proceedings of the EACL 2006 workshop on Adaptive Text Extraction and Mining
(ATEM 2006), Trento, Italy (2006) 1 – 8
5. Hearst, M.: Automatic acquisition of hyponyms from large text corpora. In: Proceedings of
the 14th conference on Computational linguistics, Nantes, France (1992) 539 – 545
6. Crescenzi, V., Mecca, G.: Automatic information extraction from large websites. Journal of
the ACM 51(5) (2004) 731 – 779
7. Downey, D., Etzioni, O., Soderland, S.: A probabilistic model of redundancy in information
extraction. In: 19th International Joint Conference on Artificial Intelligence (IJCAI’05),
Edinburgh, UK (2005) 1034 – 1041
8. Downey, D., Broadhead, M., Etzioni, O.: Locating Complex Named Entities in Web Text.
In: Proceedings of the Twentieth International Joint Conference on Artificial Intelligence
(IJCAI’07), Hyderabad, India (2007)
9. Sumida, A., Torisawa, K., Shinzato, K.: Concept-instance relation extraction from simple
noun sequences using a full-text search engine. In: Proceedings of the ISWC 2006 workshop
on Web Content Mining with Human Language Technologies (WebConMine), Athens, GA
(2006)
10. Cimiano, P., Staab, S.: Learning by Googling. SIGKDD Explorations Newsletter 6(2) (2004)
24 – 33
11. Ravichandran, D., Hovy, E.: Learning surface text patterns for a question answering system.
In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL 2002), Philadelphia, PA (2002) 41 – 47
12. Cilibrasi, R., Vitanyi, P.:
Automatic meaning discovery using Google.
http://www.cwi.nl/˜paulv/papers/amdug.pdf (2004)
13. Zadel, M., Fujinaga, I.: Web services for music information retrieval. In: Proceedings of
5th International Conference on Music Information Retrieval (ISMIR’04), Barcelona, Spain
(2004)
14. Véronis, J.: Weblog (2006) http://aixtal.blogspot.com.
15. Geleijnse, G., Korst, J., de Boer, V.: Instance classification using co-occurrences on the web.
In: Proceedings of the ISWC 2006 workshop on Web Content Mining with Human Language
Technologies (WebConMine), Athens, GA (2006) http://orestes.ii.uam.es/workshop/3.pdf.
16. Mori, J., Tsujishita, T., Matsuo, Y., Ishizuka, M.: Extracting relations in social networks from
the web using similarity between collective contexts. In: Proceedings of the 5th International
Semantic Web Conference (ISWC 2006). Volume 4273 of LNCS., Athens, GA, Springer
(2006) 487 – 500
17. Jin, Y., Matsuo, Y., Ishizuka, M.: Extracting a social network among entities by web mining.
In: Proceedings of the ISWC 2006 workshop on Web Content Mining with Human Language
Technologies (WebConMine), Athens, GA (2006)
18. Zhou, G., Su, J.: Named entity recognition using an hmm-based chunk tagger. In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL
2002), Philadelphia, PA (2002) 473 – 480
19. Brothwick, A.: A Maximum Entropy Approach to Named Entity Recognition. PhD thesis,
New York University (1999)
20. Finkel, J.R., Grenager, T., Manning, C.D.: Incorporating non-local information into information extraction systems by gibbs sampling. In: Proceedings of the 43rd Annual Meeting
of the Association for Computational Linguistics (ACL 2005), Ann Arbor ,MI (2005)
21. Korst, J., Geleijnse, G., de Jong, N., Verschoor, M.: Ontology-based extraction of information from the World Wide Web. In: Intelligent Algorithms in Ambient and Biomedical
Computing. Philips Research Book Series. Springer (2006) 149 – 167